Hello Everybody,
I have few queries :
Do you have any idea how to check the performance of HPC cluster having mpich on top of centos 6.2?
Are there any standard programs (like FFT, graphics rendering etc) to check the performance of mpi cluster with single node and multiple nodes?
Can we check the performance of single core and dual core?
Can we benchmark small cluster of 4 nodes or is benchmarking softwares are only for big clusters?
Thanks..
the standard measuring tool is HPL which solves for a very large multidimensional array. Prepackaged rpms might be found in the Rocks distribution. The answer is yes to each of your other questions
Thanks for your help Otheus.
I am new to the field of HPC. I installed HPCC and HPL. Even I am able run it and get the results. But I am not able to understand it. Also its running for my colete cluster, I also want to run them for my single machine. And now I am not able to tell whether its running on both the cores of my machine or only one process per machine, as I am having core 2 duo machines.
"Understanding" isn't so important. At the end of the results, there is a number labeled "GLOPS". It looks like this:
WR00L2C2 8192 16 1 1 327.60 1.119e+00
The first column expresses some of the options used in the options (HPL.dat) file. The next two columns express the size in dimensions. The last column -- the one you want -- is the number of Gigaflops measured for that run. It's in scientific notation, meaning:
1.119 x 10^(00) = 1.119 GFlops
If the output were 3.42e-01, then you'd have:
3.42 x 10^(-01) = 0.342 GFlops = 342 MFlops
Do you understand now?
---------- Post updated at 10:02 AM ---------- Previous update was at 09:34 AM ----------
As to how to configure this to run on multiple nodes, you have to (1) configure HPL.dat to run on mxn processors, where m and n are close together; so if you have a cluster containing 12 cores, optimal is normally m=3 and n=4. That's not always the case, but it typically is. (2) You must configure your MPI environment so that it runs on that many processors. How you do #2 depends on your MPICH software and setup. If you have a batch system like PBS or SGE, that further complicates the matter, as you need to queue the job with that number of processors. Sometimes with queuing engines, you can't easily dictate which nodes the job will run on. So if you have 6 nodes each with 4 cores, it might not be possible to say "run this on 3 nodes using all cores on each one". (But normally it's possible to say "run it on 3 nodes using 1 core on each one".)
Thanks Otheus for your generous reply.
Its very helpful.
I configured my system and somehow able to get the results from HPL.
I used one website which says "How do I tune my HPL.dat file?" http: //www.advancedclustering. com/faq/how-do-i-tune-my-hpldat-file.html to create the auto generated HPL.dat file.
The parameters which I used are :
Nodes: 5
Cores per Node: 2
Memory per Node (MB): 2048
Block Size (NB): 128
But I don't know whether the parameters which I used are correct or not.
I built a cluster of 5 PC's and all of them are having intel core 2 duo processors. I don't even know what this block size signifies.
Also, is there a way to chaeck the performance of single PC with respect to its individual core? As I am having Core 2 duo PC, I want to check the performance of this PC while its using single core and on the other hand when it uses both cores.
Thanks.
---------- Post updated at 01:46 PM ---------- Previous update was at 01:30 PM ----------
I have one more query.
While going through website http ://hpl-calculator.sourceforge.net/
I came across one more parameter Operations Per Cycle, which by default they used 4. I don't understand how to calculate this value. I googled and there are various terms, flops, mips, instructions per cycle, instructions per second and other things. can you tell me how to calculate this value for core 2 duo processor and i7 processor? As now a days multiple cores are there in PCs this per second calculations are for single core or multiple core?
Thanks again...